Pantograph state evaluation and fault diagnosis integrated method and system and storage medium

Abstract

The application discloses a pantograph-catenary state evaluation and fault diagnosis integrated method and system and a storage medium, and belongs to the technical field of intelligent monitoring of pantograph-catenary of rail transit. The application firstly pre-processes pantograph-catenary detection data; generates an offline state sample set through engineering rule priori, fusion sample graph construction, weak supervision state propagation and label correction; trains an Ordinal SVM model using the offline state sample set for online evaluation of the pantograph-catenary state; then constructs an improved TabM model with a fusion prototype constraint and a class balance mechanism, and realizes accurate small sample fault diagnosis after Bayesian optimization; and finally, joint discrimination is performed through state-fault consistency constraints, and a joint discrimination result is output. The application solves the pain points of missing pantograph-catenary state labels, insufficient use of ordered information, low small sample fault diagnosis precision and result conflicts, and the detection result has strong engineering interpretability and can be widely applied to online operation and maintenance scenes of the pantograph-catenary of rail transit.

Description

Technical Field

[0001] This invention relates to the field of pantograph-catenary operation monitoring and intelligent fault diagnosis technology in rail transit, and in particular to an integrated method, system and storage medium for pantograph-catenary state assessment and fault diagnosis based on weakly supervised learning and ordinal support vector machine. Background Technology

[0002] The pantograph-catenary system, consisting of the pantograph and the overhead contact line, is a core component of rail transit traction power supply. Its operational status directly affects current collection quality, power supply continuity, and train operation safety. When pantograph-catenary system malfunctions or localized faults are not identified in a timely manner, it can easily lead to power outages, train delays, or even major operational accidents. With the continuous application of online pantograph-catenary monitoring technology, the ability to acquire status information during pantograph-catenary operation has been continuously enhanced, providing a foundation for condition assessment and fault diagnosis based on monitoring data.

[0003] In recent years, research on pantograph-catenary condition monitoring has deepened, and existing studies can be divided into two categories. One category focuses on fault or defect detection, primarily targeting the identification of anomalies in arcing, insulators, support devices, pantograph contactors, and local components. This type of research has improved the automatic identification capability of local anomalies or typical defects, but most methods still focus on single fault phenomena or local components, making it difficult to directly reflect the overall operating status of the pantograph-catenary system. The other category focuses on operating status or risk analysis, attempting to characterize the system's operating level through probabilistic risk modeling, reliability analysis, or critical state quantity estimation. This type of research can characterize the operating status of the pantograph-catenary or overhead contact system at the system level, but it typically focuses more on risk levels, fault propagation relationships, or critical state quantity estimation, and has not yet fully addressed the issue of consistent expression between condition assessment results and specific fault types in online monitoring scenarios.

[0004] Therefore, existing research provides an important foundation for the utilization of pantograph-catenary detection data, but it still has shortcomings in online operation and maintenance scenarios. First, pantograph-catenary online detection data has obvious multi-source heterogeneous characteristics. Different detection quantities have significant differences in physical meaning, dimensional scale, and anomaly sensitivity. However, many studies still mainly focus on single-class features or single-class faults, and have not yet formed a unified organizational method for the overall state of the pantograph-catenary system and typical faults. Second, state assessment and fault diagnosis are often handled separately in existing studies. Fault diagnosis can usually construct labels based on rules or human experience and carry out supervised learning, while state assessment often relies on offline methods such as scoring, weighting, and clustering, which are difficult to directly serve online identification. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies, such as missing pantograph-catenary state labels, insufficient utilization of ordered level information, low accuracy of fault diagnosis with small samples, and lack of consistency between state and fault outputs. This invention provides an integrated method for pantograph-catenary state assessment and fault diagnosis. It constructs a reliable ordered state dataset without manual state labels, utilizes ordinal support vector machines to achieve accurate ordered state assessment, combines an improved small-sample table model to enhance fault diagnosis performance, and achieves coordinated output of state and fault through consistency constraints, thereby improving the accuracy and engineering interpretability of online pantograph-catenary detection.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The integrated method for pantograph-catenary condition assessment and fault diagnosis includes the following steps: S1. The vehicle-mounted pantograph-catenary detection system collects comprehensive data samples at positioning points along the railway line to construct a positioning point-level dataset. The comprehensive data includes pantograph-catenary geometric characteristics, dynamic contact characteristics, electrical and thermal characteristics, and positioning condition information. S2. Preprocess the raw data of the location point-level dataset, including sample validity screening, feature field screening, continuous feature standardization, and rearranging the samples according to the location point order. S3. Establish a pantograph-catenary condition assessment index system and define the pantograph-catenary condition as three levels: healthy, early warning, and fault. S4. Based on the rules of pantograph-catenary maintenance engineering, the initial state prior of the sample is generated, and combined with the spatial-temporal-feature fusion sample map, weakly supervised state propagation and label correction are performed on the positioning point sample to generate an offline state sample set. S5. Train the Ordinal SVM model using the offline state sample set as supervised information; the vehicle-mounted pantograph-catenary detection system collects the latest sample to be detected, inputs it into the trained Ordinal SVM model, and outputs the corresponding state level; S6. Construct an improved TabM model that integrates prototype constraint mechanism and class balance mechanism, and use Bayesian optimization to globally optimize the model hyperparameters; input the sample to be detected into the improved TabM model constructed by the optimal parameters, and output the fault type; the fault type includes health status, abnormal guide height, abnormal pull-out value, abnormal slope value, abnormal arcing, and abnormal hard point; S7. Under the consistency constraint of the status assessment results and fault diagnosis results, when the two types of outputs meet the consistency constraint, the original results are directly output; when the two types of outputs conflict, they are corrected or marked for verification according to the matching relationship between the status level and the fault type; the final status level and fault type of the location point sample are output to form a joint discrimination result.

[0007] Furthermore, step S4 specifically includes the following steps: S41. Based on the over-limit classification rules of pantograph-catenary maintenance engineering, all samples are judged by single index rules and combined rules to generate initial state priors. Samples that meet the explicit rule conditions are directly assigned a definite state label, and samples that are in the critical interval or have multiple rule conflicts are assigned a weak label prior. S42. Construct a spatial correlation graph representing the continuity between adjacent samples within the same detection sequence, a temporal correlation graph representing the consistency between samples from similar locations in the same station area across different detection files, and a feature correlation graph representing the similarity between samples with similar state patterns. Weighted fusion of these three types of graphs yields a unified sample graph. Based on this unified sample graph, a normalized label propagation matrix is ​​constructed. ; ; in, For the normalized label propagation matrix, To unify the association weight matrix of sample graphs, For degree matrix, For matrix Element; For matrix The elements represent samples to sample Weighted edges; To unify the number of nodes in the sample graph, i.e., the total number of samples after preprocessing; S43. Using high-confidence state-labeled samples as anchors, iteratively correct the state responses of weak-labeled samples based on the normalized label propagation matrix. After convergence, the final state labels of all samples are obtained, and the samples with the final labels form an offline state sample set. The iterative formula for label propagation is: ; in, Indicates the first The state response matrix of the next iteration For the first The state response matrix of the next iteration This is the initial soft-label prior matrix. denoted as the propagation coefficient.

[0008] Furthermore, the total loss function of the improved TabM fault diagnosis model described in step S6 is: ; in, For the total loss, To balance the classification loss and alleviate class imbalance; The prototype constraint loss is used to constrain the latent space representations of similar fault samples to cluster towards the class prototype center. These are the weighting coefficients for the prototype constraint terms; The method for calculating the category-balanced classification loss is as follows: ; in, Indicates sample The actual fault type label, Indicates sample The corresponding weight of the fault type The size is related to the number of samples corresponding to the fault type; the fewer the number of samples corresponding to a fault type, the greater the weight. Indicates sample The diagnostic probability of the corresponding fault type; The prototype constraint loss is calculated as follows: ; in, For the sample The latent representation output by the shared feature extractor Indicates sample The prototype vector corresponding to the fault type; the prototype vector is calculated as follows: ; in, Indicates the first Class prototypes of faults in the latent space , For the first A sample set of fault types. This represents the corresponding number of samples.

[0009] Furthermore, the hyperparameters for Bayesian optimization in step S6 include the number of TabM branches, network block dimension, number of network blocks, dropout rate, learning rate, weight decay coefficient, batch size, and prototype constraint weight coefficient.

[0010] Furthermore, the consistency constraint rule described in step S7 is as follows: When the status assessment result is output as healthy, the fault diagnosis result should be output as healthy accordingly. When the status assessment result output is a warning, the fault diagnosis result should output as abnormal guide height and abnormal pull-out value. When the condition assessment result outputs a fault, the fault diagnosis result should output an abnormal slope value, an abnormal arcing value, or an abnormal hard point.

[0011] The present invention also provides an integrated system for pantograph-catenary condition assessment and fault diagnosis for performing the above-described method, comprising: The data preprocessing module is used to receive the detection data collected by the pantograph-catenary inspection system and preprocess it. The weakly supervised state modeling module is used for engineering rule priors, construction of space-time-feature fusion sample graphs, and weakly supervised state propagation and label correction. The ordered state evaluation module has a built-in trained OrdinalSVM model, which is used to output the state level of the sample to be detected. The small sample fault diagnosis module has a built-in improved TabM model that has been trained. The improved TabM model integrates prototype constraint mechanism and class balance mechanism and is optimized and tuned by Bayesian method to output the fault type of the sample to be detected. The joint discrimination output module is used to perform consistency verification and conflict correction on the state level output by the OrdinalSVM model and the fault type output by the improved TabM model based on preset consistency constraints, and output the joint discrimination result at the location point level.

[0012] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described integrated method for pantograph-catenary condition assessment and fault diagnosis.

[0013] Compared with the prior art, the present invention has the following advantages: (1) Realize reliable ordered state modeling without manual labeling: This invention combines the label propagation algorithm of engineering rule prior and spatiotemporal feature fusion graph to construct an ordered state dataset that combines engineering rationality and data distribution consistency without manual labeling of the state; it solves the pain point of missing state labels in online detection data; (2) The ordered state assessment has high accuracy and conforms to the laws of physical evolution: This invention uses Ordinal SVM for state assessment, which makes full use of the orderliness of the state levels and constrains the discrimination boundary through hierarchical binary classification to avoid cross-level misjudgment of ordinary multi-classification models; Experimental results show that the state assessment accuracy reaches 0.9733 and the AUC reaches 0.9912. Misjudgment only occurs between adjacent levels, which is highly consistent with the physical characteristics of the gradual degradation of the bow-net state; (3) Excellent fault diagnosis performance in small sample imbalance scenarios: This invention improves the TabM model by introducing a prototype constraint mechanism to enhance the compactness of the representation of similar faults, introducing a class balance mechanism to alleviate the sample distribution bias, and achieving the global optimal configuration of hyperparameters through Bayesian optimization, which effectively improves the identification accuracy of minority class faults. (4) The integrated output has strong engineering interpretability: The present invention constructs a state-fault consistency joint discrimination mechanism, constrains and corrects the matching relationship between state level and fault type, avoids the problem of contradiction between the two types of output, and the detection results are more in line with engineering cognition, which can directly support the decision-making of pantograph-catenary online operation and maintenance and preventive maintenance. Attached Figure Description

[0014] Figure 1 This is the overall flowchart of the integrated pantograph-catenary condition assessment and fault diagnosis method of the present invention; Figure 2 A schematic diagram showing the composition and data flow of a vehicle-mounted pantograph-catenary detection system; Figure 3 A comparison diagram of the state distributions obtained by offline state modeling using three methods in Example 1; Figure 4 This is a comparison of the online status assessment results indicators of the five methods in Example 2; Figure 5 A comparison of the ROC values ​​of the online status assessment results of the five methods in Example 2; Figure 6 This is the confusion matrix of the online state assessment results of the method proposed in Example 2 of this invention; Figure 7 A comparison of fault diagnosis result indicators of the five methods in Example 3; Figure 8 A comparison of ROC and PR curves of the fault diagnosis results of the five methods in Example 3; Figure 9 This is the confusion matrix of the fault diagnosis results of the method proposed in Example 3. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] I. Methods, System and Storage System The flowchart of an integrated method for pantograph-catenary condition assessment and fault diagnosis can be referred to... Figure 1 The specific steps are as follows.

[0017] 1. Acquisition of pantograph-catenary line inspection data The pantograph-catenary line detection data is acquired online by the vehicle-mounted pantograph-catenary line detection system. The composition and data flow of the vehicle-mounted pantograph-catenary line detection system are as follows: Figure 2As shown, the vehicle-mounted pantograph-catenary inspection system consists of a roof section, an interior section, and an undercarriage section. The roof section includes a camera unit and a mechanical data acquisition unit. The camera unit acquires video monitoring, arcing, temperature, and geometric and wear detection information for the pantograph-catenary system. The mechanical data acquisition unit acquires hard point and pressure measurement information. The interior section mainly includes an in-vehicle processing host and a fiber optic data analyzer, responsible for receiving, processing, and summarizing data from each detection module. The undercarriage section includes a vibration compensation device to compensate for attitude disturbances during vehicle operation. During data transmission, the multi-source detection information collected by the roof-mounted camera unit, the mechanical data acquisition unit, and the undercarriage vibration compensation device is uniformly transmitted to the in-vehicle processing host for further data processing and storage, providing the initial foundation for subsequent data organization and analysis.

[0018] Sample set construction: Data is recorded at each positioning point along the track to form a positioning point-level sample set. Each sample data corresponds to the comprehensive detection data at the positioning point during train operation, including track positioning, operating conditions, catenary geometric parameters, pantograph-catenary dynamic parameters, and electrical and thermal related parameters. The data composition is shown in Table 1. Table 1 Composition of pantograph-catenary line inspection data .

[0019] 2. Data Preprocessing (1) Sample cleaning (validity screening) If a sample has missing key fields, or incomplete location or time information, it is considered an invalid record and is removed. The validity variables for each sample are: ;in, For the first The original feature vectors corresponding to each sample can then be represented as follows: ; in, Indicates the first Location point identification information corresponding to each sample Indicates the first The detection time corresponding to each sample.

[0020] (2) Sample feature field screening like ; After removing redundant fields that have no direct correlation with subsequent evaluations, have limited information content, or mainly serve an indexing function, the first... The feature vector of each sample becomes: ; in, The number of features before filtering the field. The number of features to retain after filtering a field.

[0021] (3) Feature standardization Because the dimensions of different detection quantities differ significantly, the value ranges of related variables are not consistent. Therefore, a unified scaling process is applied to continuous variables. For any continuous feature... Standardized processing is adopted: ; in, and They represent the first The mean and standard deviation of each feature on the sample set. After processing, the... The feature vector of a sample can be written as: .

[0022] (4) Sample space sorting Considering that the pantograph-catenary line detection data is continuously collected along the line and that the samples have a clear spatial order, the samples are further rearranged according to the order of the positioning points. Let the rearrangement operator be... Then the final sample sequence can be represented as: ; in, This represents the number of valid samples after sample cleaning. The positioning point sequence must be consistent.

[0023] After the above processing, the original online detection records are organized into a sample set with complete fields, uniform scale, and maintained spatial order, providing a unified input for subsequent evaluation and diagnosis.

[0024] 3. Pantograph-Catenary System Condition Assessment Method (1) Status assessment index system and level classification The condition assessment indicators are divided into three categories: geometric condition indicators, dynamic contact indicators, and electrical operation indicators, and the corresponding sample representations are as follows: ; in, , , These represent three subsets of indicators: geometric state, dynamic contact, and electrical operating state. For specific indicators, please refer to Table 1.

[0025] Considering that the pantograph-catenary state evolves gradually from normal to abnormal, the pantograph-catenary state set is defined as a three-level ordered set: , ; in, , , These correspond to three status levels: health, warning, and fault. The definitions of each level are shown in Table 2. Table 2 Classification of Pantograph-Catenary System Condition Levels .

[0026] (2) Weakly supervised state modeling based on rule guidance and spatiotemporal feature correlation (2.1) State priors based on engineering rules: The original online detection data lacks complete manual status labels. This invention generates the initial state prior of the samples based on the exceedance classification rules in the on-site management documents, combined with combination rules. The corresponding rules are shown in Table 3: Table 3 Prior Rules for Initial Sample State ; For samples that meet the explicit rule conditions, a definite state label is directly assigned; for samples that are in the critical interval or have multiple rule conflicts, only a weak prior label is assigned for subsequent spatiotemporal correlation correction.

[0027] (2.2) Construction of sample graph based on multi-source association: Single-point rules can provide high-confidence initial state priors for some samples, but they are not stable enough for boundary samples and samples with conflicting rules. Therefore, a sample graph is further constructed to characterize the relationships between location point samples. The sample graph can be represented as follows: , where the set of nodes Corresponding location point samples, edge set This indicates the adjacency relationship between samples. Let be the edge weight matrix.

[0028] Considering that online pantograph-catenary inspection data has characteristics such as continuous acquisition along the line, repeated observation across inspection cycles, and similar local state patterns, this invention constructs spatial correlation graphs respectively. Time Relationship Diagram Feature association graph Among them, the spatial map is used to describe the local continuity between adjacent points within the same detection sequence; the temporal map is used to describe the consistency of samples from the same station area and similar locations between different detection files; and the feature map is used to describe the similarity between samples with similar state patterns.

[0029] Based on this, the three types of graphs are linearly fused according to their weights to obtain a unified sample graph: , ; in, , and These represent the fusion weights corresponding to the spatial correlation graph, temporal correlation graph, and feature correlation graph, respectively. Therefore, the matrix... Each element This comprehensively reflects the sample and samples Overall correlation strength in terms of spatial location, detection timing, and feature patterns; Further define the degree matrix : ; For matrix Element; And construct the normalized propagation matrix: ; Among them, matrix Used to describe the structure of the original uniform sample graph, matrix This is then used for subsequent state propagation calculations. Normalization reduces the impact of differences in node connectivity on the propagation results, making label propagation more stable.

[0030] (2.3) Weakly supervised state propagation and label correction Based on the multi-source associated sample graph, high-confidence rule samples are used as label anchors. Under the constraints of the graph structure, the state response of the remaining samples is iteratively updated to constrain the weakly supervised state propagation and correct the initial state prior.

[0031] First, healthy and faulty anchors are identified based on comprehensive risk and key rule responses. Healthy anchors correspond to samples with low overall risk; faulty anchors correspond to samples with high overall risk or those exhibiting severe anomalies in multiple key indicators simultaneously. Anchor samples retain their labels during propagation. Then, an initial soft label matrix is ​​constructed based on the comprehensive risk: ,in, This represents the number of status tags. To enhance the discriminative power of the initial labels, state prototypes are constructed using health and failure anchors, and the soft labels are then corrected based on the distance from the sample to the prototype. Let the first The state response matrix for the next iteration is: Then its update method is: ,in, The propagation coefficient; After each iteration, label constraints are reapplied to the anchor point samples, and... Each row is normalized; After iterative convergence, initial state labels are first obtained based on comprehensive risk and classification thresholds. Then, the adjacent level responses obtained through propagation are used to locally correct the boundary samples, while retaining the original labels for healthy and faulty anchor points. Thus, the final state labels retain both the engineering basis of rule classification and the local consistency of graph propagation.

[0032] (2.4) Construction of offline sample set After initial state prior generation, fusion sample graph construction, and weakly supervised state propagation and label correction, the final state label for each sample can be obtained. Let the sample... The final state label is denoted as The offline state sample set can then be represented as: ; in, These correspond to three state levels: health, warning, and fault. Based on this, an offline sample set with a unified state label representation was constructed. This sample set, while retaining the original detection features, introduces state labels based on rule constraints and graph propagation correction, which can serve as the training basis for subsequent online state assessment models.

[0033] (3) State assessment based on Ordinal SVM model Considering the explicit sequential relationship between the pantograph-catenary state labels, this invention employs Ordinal SVM to construct an online state assessment model. Unlike ordinary multi-classification methods, Ordinal SVM can utilize the ordered nature between state levels, making it more suitable for pantograph-catenary state recognition tasks.

[0034] To utilize state sequence information, this invention transforms the three-level state classification into... The ordered binary classification subproblem. For the... For each sub-problem, the binary label is defined as: , ; Thus, the original 3-class classification problem is decomposed into two binary classification tasks, corresponding to whether it is above healthy, whether it is above good, and whether it is above warning.

[0035] For each subproblem, establish the corresponding SVM discriminant function: ; in, Represents the sample feature mapping, and The first The parameters of each sub-classifier are used. During training, generated offline state samples are used as supervision information to learn two levels of discrimination boundaries.

[0036] For any online input sample Its final state level is determined jointly by the discrimination results of each level: ; in, The indicator function is denoted by . As a sample successively crosses more levels of discrimination boundaries, its corresponding state level also increases. Therefore, the Ordinal SVM model can output three state results. Compared to ordinary multi-class SVM, Ordinal SVM no longer treats each state as an independent, unordered category, but maintains the evolutionary relationship between state levels through hierarchical discrimination, consistent with the characteristic of the pantograph-catenary system gradually developing from normal operation to abnormal failure. Therefore, it is more suitable for online state assessment scenarios.

[0037] 4. TabM-based method for small-sample pantograph-catenary fault diagnosis Based on the TabM model, a prototype constraint mechanism and a class balance mechanism are introduced to construct a small-sample fault diagnosis model. BO is used to optimize the model hyperparameters (and Bayesian optimization) to improve fault diagnosis accuracy and generalization performance.

[0038] (4.1) Basics of the TabM Model Let the input sample be ,in, Let be the dimension of the input features. First, the original input is mapped to the latent space using a shared feature extractor, resulting in a shared representation of the samples: ; in, Indicates by parameters Shared feature mapping function of control, Let be the representation vector of the sample in the latent space. The dimension is represented by the latent space. In the shared representation... Based on this, build The first branch. For the first... Branches: ; in, Indicates the first Mapping function for each branch, For the corresponding branch parameters, This is the implicit representation of the output of this branch. Furthermore, the... The diagnostic probability of each branch for each fault type can be expressed as: ; in, and The first The weight matrix and bias vector of each branch output layer , This represents the number of fault types.

[0039] After obtaining the output results of each branch, the final diagnostic probability is obtained by means aggregation: ; Therefore, the model ultimately outputs the diagnostic probability distribution of samples for each fault type, and the diagnostic results are as follows: ; in, Represents the final output vector The middle sample belongs to the first The probability of diagnosing this type of fault.

[0040] For the training sample set, the basic TabM model uses cross-entropy loss for supervised learning, and its classification loss can be expressed as: ; in, For the sample In the The actual label indication quantity on the fault type, This represents the diagnostic probability of the fault type corresponding to the model. By minimizing the above loss, the model can learn the mapping relationship between input features and fault types.

[0041] (4.2) Prototype constraint mechanism To enhance the clustering of similar fault samples, this invention introduces a prototype constraint mechanism based on the TabM shared feature representation.

[0042] Let the sample The latent representation after the shared feature extractor is , No. The sample set corresponding to the fault class is denoted as The sample size is Then the first The category prototypes for fault classes are defined as follows: ; in, Indicates the first The class fault is centrally represented in the latent space. Based on this, to constrain samples of the same class to converge towards their class prototype, the prototype constraint loss is defined as follows: ; in, Indicates sample The prototype vector corresponding to the fault category. This constraint improves the intra-class compactness of fault samples of the same type in the feature space by reducing the distance between the latent representation of the sample and the prototype of the same class.

[0043] For small-sample fault data of the pantograph-catenary system, prototype constraints can use class centers to impose structural constraints on finite samples, thereby reducing the impact of individual sample fluctuations on class representation and enhancing the clustering of minority class faults in the latent space, thus improving the stability of fault representation under small-sample conditions.

[0044] (4.3) Category Balance Mechanism To mitigate the impact of uneven category distribution on diagnostic performance, this invention further introduces a category balancing mechanism, assigning differentiated weights to different fault types.

[0045] Let the first The number of samples for this type of fault is The corresponding category weight is denoted as The weights are related to the number of samples in each class; fault types with fewer samples receive larger weights to enhance the contribution of minority class samples during model training. Based on this, the class-balanced classification loss is as follows: ; in, Indicates sample The corresponding weight of the fault type Indicates sample The diagnostic probability of the fault type.

[0046] By using the weighting method described above, the influence of minority class fault samples in the loss function is appropriately enhanced, thereby weakening the dominant role of majority class samples in model training and improving the training balance between different fault categories.

[0047] After introducing prototype constraints and class balancing mechanisms, the model's total loss function can be expressed as follows: ; in, The weights are the prototype constraint terms. This loss function simultaneously considers fault classification performance, minority class sample learning, and intra-class representation compactness, providing a unified optimization objective for training small-sample pantograph-catenary fault diagnosis models.

[0048] (4.4) BO-based hyperparameter optimization The diagnostic performance of the improved TabM model is highly sensitive to hyperparameter configuration, and there are significant coupling relationships between different parameters, making it difficult to obtain a stable optimal combination through manual parameter tuning. Therefore, Boolean search (BO) is used to perform a unified search for key hyperparameters of the model to improve its performance in small-sample fault diagnosis.

[0049] Let the hyperparameter vector be Its search space is denoted as The hyperparameter optimization problem of this invention can then be expressed as: ; in, Indicates parameter combination The performance metrics of the model on the validation set are shown below. This invention incorporates all relevant parameters into the BO search space and uses the validation set macro-F1 as the optimization objective. The search space is shown in Table 4: Table 4 Hyperparameter Definitions and Search Space .

[0050] Furthermore, this invention employs a TPE-based sampling strategy to perform parameter search. For each trial, a set of candidate hyperparameters is first generated from the current trial, and an improved TabM model is constructed accordingly; subsequently, model training, validation set performance evaluation, and optimal parameter updates are completed. The number of BO trial rounds is set to 50, and early stopping is used to complete model selection in a single trial.

[0051] In each BO trial, the training objective of improving TabM is: ; Therefore, BO not only optimizes the structural and training parameters of TabM, but also optimizes the prototype constraint strength, thereby achieving unified coordination between basic classification, class balance and prototype constraints.

[0052] Let the first The parameter combination corresponding to this experiment is: Then its target value can be expressed as: ; After multiple iterations, BO outputs the optimal parameter combination. And based on this, a model is constructed.

[0053] 5. Joint judgment of condition assessment and fault diagnosis The outputs of pantograph-catenary condition assessment and fault diagnosis are not independent. The condition level characterizes the overall operational status of the pantograph-catenary system, while the fault type reflects the source of the anomaly and the dominant defect characteristics; both should be consistent. Therefore, this invention constructs a joint discrimination framework for condition assessment and fault diagnosis, using consistency constraints and joint decision-making rules to uniformly correct and output the two types of results, forming a joint discrimination result at the location point level.

[0054] (5.1) State-fault consistency constraint mechanism The condition assessment results and fault diagnosis results respectively characterize the overall level of pantograph-catenary operation and the source of anomalies, and the two should be consistent. For ease of implementation, this invention further uses a tabular format to constrain typical consistency relationships, as shown in Table 5: Table 5 Condition-Fault Consistency Relationship .

[0055] (5.2) Joint Output and Decision Rules Based on consistency constraints, this invention further constructs a joint decision function to provide a unified output for both the state assessment results and the fault diagnosis results. Let the output of the state assessment branch be... The fault diagnosis branch output is The joint output can then be expressed as: ; in, Denotes the joint decision function. This refers to the final state level and fault type.

[0056] When both types of output satisfy the consistency constraint, the original result is output directly; when the two types of output conflict, correction or verification is performed based on the matching relationship between the state level and the fault type. Specific rules are shown in Table 6. Table 6 Joint Amendment Rules .

[0057] An integrated system for pantograph-catenary condition assessment and fault diagnosis for performing the above method includes: The data preprocessing module is used to receive the detection data collected by the pantograph-catenary inspection system and preprocess it. The weakly supervised state modeling module is used for engineering rule priors, construction of space-time-feature fusion sample graphs, and weakly supervised state propagation and label correction. The ordered state evaluation module has a built-in trained OrdinalSVM model, which is used to output the state level of the sample to be detected. The small sample fault diagnosis module has a built-in improved TabM model that has been trained. The improved TabM model integrates prototype constraint mechanism and class balance mechanism and is optimized and tuned by Bayesian method to output the fault type of the sample to be detected. The joint discrimination output module is used to perform consistency verification and conflict correction on the state level output by the OrdinalSVM model and the fault type output by the improved TabM model based on preset consistency constraints, and output the joint discrimination result at the location point level.

[0058] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described integrated method for pantograph-catenary condition assessment and fault diagnosis.

[0059] II. Experimental Verification Example 1: Offline Experimental Verification To verify the effectiveness of the offline state modeling method proposed in this invention, three methods—AHP+TOPSIS, K-meansClustering, and the method provided in this invention (Proposed)—were compared. Their distribution results across the three states were statistically analyzed. The experimental results are as follows: Figure 3 As shown in Table 7.

[0060] Table 7 Comparison Results of State Distribution ; The results show significant differences in the state distributions obtained by different methods. The AHP+TOPSIS results are mainly concentrated in L3, while the K-means Clustering results are mainly distributed in L1 and L2, failing to effectively identify L3. This indicates that when relying solely on weighted comprehensive evaluation or pure unsupervised clustering, the state partitioning results are easily affected by the evaluation rule settings or sample distribution characteristics, making it difficult to simultaneously consider both engineering priors and sample structure information.

[0061] In contrast, the state distribution obtained by the method (Proposed) provided in this invention exhibits a more clearly defined hierarchy. Most samples are classified as L1, while a small number are identified as L2 and L3, consistent with the actual patterns in online pantograph-catenary detection scenarios. Furthermore, when comparing and reviewing fault data, the Proposed method achieves a 100% consistency rate in identifying L3, demonstrating that the proposed state modeling method effectively preserves the significant anomalous features of fault samples and reliably identifies key risk states during offline label construction. This also effectively proves that the offline state results not only have better engineering interpretability but also provide a more reliable supervised label foundation for subsequent online state assessment.

[0062] Example 2: Online Evaluation and Verification After obtaining the offline state labels, online state evaluation experiments were further conducted based on the constructed offline state sample set. It is worth noting that, to balance the sample size, only 600 sets of L1 data were selected. The proportions of the training set, validation set, and test set are shown in Table 8. This invention selects five methods for comparison: SVM, Hist GBDT, FT Transformer, Tab Transformer, and the method provided in this invention (Proposed). The overall results are as follows: Figure 4 As shown in Table 8.

[0063] Table 8 Comparison of Evaluation Result Indicators ; Figure 4The results show that, overall, Proposed, SVM, and Hist GBDT significantly outperform FT Transformer and Tab Transformer, with Proposed showing the best overall performance.

[0064] As shown in Table 8, the Proposed method achieved the highest values ​​in the three key metrics of Accuracy, Macro-F1, and Weighted-F1, reaching 0.963333, 0.933581, and 0.963040, respectively. The combined performance of SVM and Hist GBDT was second best, but still lower than that of the Proposed method overall. The performance of FT Transformer further declined, while Tab Transformer performed the worst in the current task. These results demonstrate that the proposed method not only has higher overall classification accuracy but also maintains better balanced recognition capabilities across different state categories.

[0065] Based on this, Figure 5 Furthermore, the ROC comparison results for the five methods are presented. Figure 5 As can be seen, the Proposed method exhibits the best overall ROC curve, with an AUC of 0.9912, the highest among all methods. The SVM method has an AUC of 0.9905, close to that of the Proposed method. The Hist GBDT and FT Transformer methods have AUCs of 0.9797 and 0.9658, respectively, while the TabTransformer method only has an AUC of 0.6676. These results demonstrate that the proposed method also possesses advantages in overall discriminative ability.

[0066] To further analyze the classification effect of the proposed method, Figure 6 The proposed confusion matrix is ​​presented. It can be seen that both L1 and L2 classes achieve high recognition accuracy, and although L3 has fewer samples, it still achieves effective recognition. The misclassifications are mainly concentrated between adjacent states, without significant cross-level misclassifications, which is consistent with the ordered nature of state assessment problems. Overall, the proposed method demonstrates superior overall performance and good classification stability in online state assessment tasks.

[0067] Example 3: Fault Diagnosis Experiment Verification To verify the effectiveness of the small-sample fault diagnosis method proposed in this invention, five methods were selected for comparison: SVM, XG Boost, CatBoost, SS Transformer, and the method provided in this invention (Proposed). The experimental results are as follows: Figure 7 As shown in Table 9.

[0068] from Figure 7As shown, the proposed method outperforms the other comparative methods in all metrics, indicating that the proposed method has the best overall performance in the current small-sample fault diagnosis task.

[0069] Table 9 Comparison of Fault Diagnosis Results ; Table 9 shows that the proposed method achieves the best results in key metrics such as Accuracy, Macro-F1, and Weighted-F1, indicating that it not only improves the overall classification accuracy but also maintains a better balanced recognition capability across different fault categories. Compared with machine learning methods and deep learning models, the proposed method consistently improves overall performance. This demonstrates that, for small-sample fault diagnosis scenarios, introducing prototype constraints and class balancing mechanisms can more effectively enhance the fault representation capability and mitigate the impact of class imbalance on training results.

[0070] Furthermore, Figure 8 The paper presents comparisons of ROC and PR curves for five methods. It can be seen that the proposed method achieves the best results under both curve types, with an AUC of 0.9686 for ROC and an AP of 0.8640 for PR, both higher than the other comparative methods. These results demonstrate that the proposed method not only performs best on discrete classification metrics but also exhibits advantages in overall discriminative ability and stable recognition under limited sample conditions. This further validates the effectiveness of the proposed improved TabM framework in multi-class fault diagnosis tasks.

[0071] Finally, we will analyze the specific classification effects of the proposed method. Figure 9 The confusion matrix of the proposed method is presented. It can be seen that the proposed method achieves relatively accurate identification for most fault categories, with misclassifications mainly concentrated between adjacent categories with similar features, and no significant large-scale misclassifications. This indicates that the model can effectively capture the main fault modes, and the remaining error mainly comes from the boundary overlap between some categories under conditions of few samples. Combined with the overall results, it can be seen that the proposed method not only has superior overall performance in small-sample fault diagnosis tasks, but also good classification stability and engineering applicability.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An integrated method for pantograph-catenary condition assessment and fault diagnosis, characterized in that, Includes the following steps: S1. The vehicle-mounted pantograph-catenary detection system collects comprehensive data samples at positioning points along the railway line to construct a positioning point-level dataset. The comprehensive data includes pantograph-catenary geometric characteristics, dynamic contact characteristics, electrical and thermal characteristics, and positioning condition information. S2. Preprocess the raw data of the location point-level dataset, including sample validity screening, feature field screening, continuous feature standardization, and rearranging the samples according to the location point order. S3. Establish a pantograph-catenary condition assessment index system and define the pantograph-catenary condition as three levels: healthy, early warning, and fault. S4. Based on the rules of pantograph-catenary maintenance engineering, the initial state prior of the sample is generated, and combined with the spatial-temporal-feature fusion sample map, weakly supervised state propagation and label correction are performed on the positioning point sample to generate an offline state sample set. S5. Train the Ordinal SVM model using the offline state sample set as supervised information; the vehicle-mounted pantograph-catenary detection system collects the latest sample to be detected, inputs it into the trained Ordinal SVM model, and outputs the corresponding state level; S6. Construct an improved TabM model that integrates prototype constraint mechanism and class balance mechanism, and use Bayesian optimization to globally optimize the model hyperparameters; input the sample to be detected into the improved TabM model constructed by the optimal parameters, and output the fault type; the fault type includes health status, abnormal guide height, abnormal pull-out value, abnormal slope value, abnormal arcing, and abnormal hard point; S7. Under the consistency constraint of the status assessment results and fault diagnosis results, when the two types of outputs meet the consistency constraint, the original results are directly output; when the two types of outputs conflict, they are corrected or marked for verification according to the matching relationship between the status level and the fault type; the final status level and fault type of the location point sample are output to form a joint discrimination result.

2. The integrated method for pantograph-catenary condition assessment and fault diagnosis according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Based on the over-limit classification rules of pantograph-catenary maintenance engineering, all samples are judged by single index rules and combined rules to generate initial state priors. Samples that meet the explicit rule conditions are directly assigned a definite state label, and samples that are in the critical interval or have multiple rule conflicts are assigned a weak label prior. S42. Construct a spatial correlation graph representing the continuity between adjacent samples within the same detection sequence, a temporal correlation graph representing the consistency between samples from similar locations in the same station area across different detection files, and a feature correlation graph representing the similarity between samples with similar state patterns. Weighted fusion of these three types of graphs yields a unified sample graph. Based on this unified sample graph, a normalized label propagation matrix is ​​constructed. ； ； in, For the normalized label propagation matrix, To unify the association weight matrix of sample graphs, For degree matrix, For matrix Element; For matrix The elements represent samples to sample Weighted edges, This represents the total number of samples after preprocessing. S43. Using high-confidence state-labeled samples as anchors, iteratively correct the state responses of weak-labeled samples based on the normalized label propagation matrix. After convergence, the final state labels of all samples are obtained, and the samples with the final labels form an offline state sample set. The iterative formula for label propagation is: ； in, Indicates the first The state response matrix of the next iteration For the first The state response matrix of the next iteration This is the initial soft-label prior matrix. denoted as the propagation coefficient.

3. The integrated method for pantograph-catenary condition assessment and fault diagnosis according to claim 1, characterized in that, The total loss function of the improved TabM fault diagnosis model described in step S6 is: ； in, For the total loss, To balance the classification loss and alleviate class imbalance; The prototype constraint loss is used to constrain the latent space representations of similar fault samples to cluster towards the class prototype center. These are the weighting coefficients for the prototype constraint terms; The method for calculating the category-balanced classification loss is as follows: ； in, Indicates sample The actual fault type label, Indicates sample The corresponding weight of the fault type The size is related to the number of samples corresponding to the fault type; the fewer the number of samples corresponding to a fault type, the greater the weight. Indicates sample The diagnostic probability of the corresponding fault type; The prototype constraint loss is calculated as follows: ； in, For the sample The latent representation output by the shared feature extractor Indicates sample The prototype vector corresponding to the fault type; the prototype vector is calculated as follows: ； in, Indicates the first Class prototypes of faults in the latent space , For the first A sample set of fault types. This represents the corresponding number of samples.

4. The integrated method for pantograph-catenary condition assessment and fault diagnosis according to claim 3, characterized in that, The hyperparameters for Bayesian optimization in step S6 include the number of TabM branches, network block dimension, number of network blocks, dropout rate, learning rate, weight decay coefficient, batch size, and prototype constraint weight coefficient.

5. The integrated method for pantograph-catenary condition assessment and fault diagnosis according to claim 1, characterized in that, The consistency constraint rule mentioned in step S7 is: When the status assessment result is output as healthy, the fault diagnosis result should be output as healthy accordingly. When the status assessment result output is a warning, the fault diagnosis result should output as abnormal guide height and abnormal pull-out value. When the condition assessment result outputs a fault, the fault diagnosis result should output an abnormal slope value, an abnormal arcing value, or an abnormal hard point.

6. An integrated system for pantograph-catenary condition assessment and fault diagnosis for performing the method as described in any one of claims 1-5, characterized in that, include: The data preprocessing module is used to receive the detection data collected by the pantograph-catenary inspection system and preprocess it. The weakly supervised state modeling module is used for engineering rule priors, construction of space-time-feature fusion sample graphs, and weakly supervised state propagation and label correction. The ordered state evaluation module has a built-in trained OrdinalSVM model, which is used to output the state level of the sample to be detected. The small sample fault diagnosis module has a built-in improved TabM model that has been trained. The improved TabM model integrates prototype constraint mechanism and class balance mechanism and is optimized and tuned by Bayesian method to output the fault type of the sample to be detected. The joint discrimination output module is used to perform consistency verification and conflict correction on the state level output by the OrdinalSVM model and the fault type output by the improved TabM model based on preset consistency constraints, and output the joint discrimination result at the location point level.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the integrated method for pantograph-catenary condition assessment and fault diagnosis as described in any one of claims 1 to 5.

Images